From 28a2d7af87ceb29b93e73c99406316c04b2c9f23 Mon Sep 17 00:00:00 2001
From: Loren Merritt <pengvado@akuvian.org>
Date: Sun, 20 Apr 2008 12:19:46 -0600
Subject: [PATCH] checkasm cabac

---
 common/cabac.c         |  4 +---
 common/cabac.h         |  9 ++++++++-
 common/x86/cabac-a.asm |  2 +-
 encoder/rdo.c          |  1 +
 tools/checkasm.c       | 33 ++++++++++++++++++++++++++++++++-
 5 files changed, 43 insertions(+), 6 deletions(-)

diff --git a/common/cabac.c b/common/cabac.c
index ad5d203f..162978e5 100644
--- a/common/cabac.c
+++ b/common/cabac.c
@@ -895,8 +895,7 @@ static inline void x264_cabac_encode_renorm( x264_cabac_t *cb )
     x264_cabac_putbyte( cb );
 }
 
-#ifndef HAVE_MMX
-void x264_cabac_encode_decision( x264_cabac_t *cb, int i_ctx, int b )
+void x264_cabac_encode_decision_c( x264_cabac_t *cb, int i_ctx, int b )
 {
     int i_state = cb->state[i_ctx];
     int i_range_lps = x264_cabac_range_lps[i_state][(cb->i_range>>6)&0x03];
@@ -909,7 +908,6 @@ void x264_cabac_encode_decision( x264_cabac_t *cb, int i_ctx, int b )
     cb->state[i_ctx] = x264_cabac_transition[i_state][b];
     x264_cabac_encode_renorm( cb );
 }
-#endif
 
 void x264_cabac_encode_bypass( x264_cabac_t *cb, int b )
 {
diff --git a/common/cabac.h b/common/cabac.h
index 1c762b89..091992fd 100644
--- a/common/cabac.h
+++ b/common/cabac.h
@@ -53,12 +53,19 @@ void x264_cabac_context_init( x264_cabac_t *cb, int i_slice_type, int i_qp, int
 
 /* encoder only: */
 void x264_cabac_encode_init ( x264_cabac_t *cb, uint8_t *p_data, uint8_t *p_end );
-void x264_cabac_encode_decision( x264_cabac_t *cb, int i_ctx, int b );
+void x264_cabac_encode_decision_c( x264_cabac_t *cb, int i_ctx, int b );
+void x264_cabac_encode_decision_asm( x264_cabac_t *cb, int i_ctx, int b );
 void x264_cabac_encode_bypass( x264_cabac_t *cb, int b );
 void x264_cabac_encode_ue_bypass( x264_cabac_t *cb, int exp_bits, int val );
 void x264_cabac_encode_terminal( x264_cabac_t *cb );
 void x264_cabac_encode_flush( x264_t *h, x264_cabac_t *cb );
 
+#ifdef HAVE_MMX
+#define x264_cabac_encode_decision x264_cabac_encode_decision_asm
+#else
+#define x264_cabac_encode_decision x264_cabac_encode_decision_c
+#endif
+
 static inline int x264_cabac_pos( x264_cabac_t *cb )
 {
     return (cb->p - cb->p_start + cb->i_bytes_outstanding) * 8 + cb->i_queue;
diff --git a/common/x86/cabac-a.asm b/common/x86/cabac-a.asm
index 26a20d08..e0c52359 100644
--- a/common/x86/cabac-a.asm
+++ b/common/x86/cabac-a.asm
@@ -81,7 +81,7 @@ endstruc
 %endif
 %endmacro
 
-cglobal x264_cabac_encode_decision, 0,7
+cglobal x264_cabac_encode_decision_asm, 0,7
     movifnidn t0d, r0m
     movifnidn t1d, r1m
     picgetgot t2
diff --git a/encoder/rdo.c b/encoder/rdo.c
index bd481cf4..8607e075 100644
--- a/encoder/rdo.c
+++ b/encoder/rdo.c
@@ -41,6 +41,7 @@ static uint16_t cabac_prefix_size[15][128];
 
 /* CABAC: not exactly the same. x264_cabac_size_decision() keeps track of
  * fractional bits, but only finite precision. */
+#undef  x264_cabac_encode_decision
 #define x264_cabac_encode_decision(c,x,v) x264_cabac_size_decision(c,x,v)
 #define x264_cabac_encode_terminal(c)     x264_cabac_size_decision(c,276,0)
 #define x264_cabac_encode_bypass(c,v)     ((c)->f8_bits_encoded += 256)
diff --git a/tools/checkasm.c b/tools/checkasm.c
index 05c25bdb..73faf122 100644
--- a/tools/checkasm.c
+++ b/tools/checkasm.c
@@ -854,6 +854,36 @@ static int check_intra( int cpu_ref, int cpu_new )
     return ret;
 }
 
+#define DECL_CABAC(cpu) \
+static void run_cabac_##cpu( uint8_t *dst )\
+{\
+    int i;\
+    x264_cabac_t cb;\
+    x264_cabac_context_init( &cb, SLICE_TYPE_P, 26, 0 );\
+    x264_cabac_encode_init( &cb, dst, dst+0xff0 );\
+    for( i=0; i<0x400; i++ )\
+        x264_cabac_encode_decision_##cpu( &cb, buf1[i]>>1, buf1[i]&1 );\
+}
+DECL_CABAC(c)
+#ifdef HAVE_MMX
+DECL_CABAC(asm)
+#else
+#define run_cabac_asm run_cabac_c
+#endif
+
+static int check_cabac( int cpu_ref, int cpu_new )
+{
+    int ret = 0, ok, used_asm = 1;
+    if( cpu_ref || run_cabac_c == run_cabac_asm)
+        return 0;
+    memcpy( buf4, buf3, 0x1000 );
+    call_c( run_cabac_c, buf3 );
+    call_a( run_cabac_asm, buf4 );
+    ok = !memcmp( buf3, buf4, 0x1000 );
+    report( "cabac :" );
+    return ret;
+}
+
 int check_all( int cpu_ref, int cpu_new )
 {
     return check_pixel( cpu_ref, cpu_new )
@@ -861,7 +891,8 @@ int check_all( int cpu_ref, int cpu_new )
          + check_mc( cpu_ref, cpu_new )
          + check_intra( cpu_ref, cpu_new )
          + check_deblock( cpu_ref, cpu_new )
-         + check_quant( cpu_ref, cpu_new );
+         + check_quant( cpu_ref, cpu_new )
+         + check_cabac( cpu_ref, cpu_new );
 }
 
 int add_flags( int *cpu_ref, int *cpu_new, int flags, const char *name )
-- 
2.39.2